Attribution Analysis Meets Model Editing: Advancing Knowledge Correction in Vision Language Models with VisEdit

TL;DR

This work studies how visual representations influence predictions in decoder-only Vision-Language Models and introduces VisEdit, a VLLM editor that leverages a two-step attribution analysis (Module Output Attribution and Visual Representation Attribution) to target edits to high-contribution visual regions. VisEdit inserts a Visual Edit Adapter (VEAD) before high-contribution layers and uses an Influence Mapper (IM) to modulate edit intensity in regions relevant to the edit prompt, training with a compound editing loss and IM loss while keeping the VLLM frozen. Empirical results on E-VQA and E-IC MMEdit across BLIP2-OPT, LLaVA-V1.5, and MiniGPT-4 show VisEdit achieving superior reliability, generality, and locality compared to adapted LLM editors, highlighting the practical viability of knowledge correction in VLLMs via targeted visual-region editing. The findings demonstrate that mid-to-late transformer layers rely on visual regions aligned with the prompt to generate responses, enabling efficient correction of factual knowledge with minimal perturbation to unrelated content.

Abstract

Model editing aims to correct outdated or erroneous knowledge in large models without costly retraining. Recent research discovered that the mid-layer representation of the subject's final token in a prompt has a strong influence on factual predictions, and developed Large Language Model (LLM) editing techniques based on this observation. However, for Vision-LLMs (VLLMs), how visual representations impact the predictions from a decoder-only language model remains largely unexplored. To the best of our knowledge, model editing for VLLMs has not been extensively studied in the literature. In this work, we employ the contribution allocation and noise perturbation methods to measure the contributions of visual representations for token predictions. Our attribution analysis shows that visual representations in mid-to-later layers that are highly relevant to the prompt contribute significantly to predictions. Based on these insights, we propose VisEdit, a novel model editor for VLLMs that effectively corrects knowledge by editing intermediate visual representations in regions important to the edit prompt. We evaluated VisEdit using multiple VLLM backbones and public VLLM editing benchmark datasets. The results show the superiority of VisEdit over the strong baselines adapted from existing state-of-the-art editors for LLMs.

Figures (9)

• Figure 1: Attribution analysis for LLaVA-V1.5 llava. E1: Measuring contributions of the attention and MLP outputs at each layer to the prediction of a key token. Average results on E-VQA MMEdit dataset are displayed in the bar chart. E2: Measuring the contributions of visual representations to the attention module outputs. The results for four samples are visualized in the heatmaps, where red indicates higher contributions and blue indicates lower. T* and L* respectively indicate the test sample index and the layer index selected by the visual representations attribution analysis. In each sample, the italicized bold text and the underlined text respectively represent the last token used for prediction and the key token to be predicted.
• Figure 2: Architecture and training loss of VisEdit.
• Figure 3: The visualization of IM module in VEAD integrated into LLaVA-V1.5. In each test, VEAD is first edited using the left image along with the prompt. Then, the outputs of IM are visualized for both images.
• Figure 4: Visualization of visual representation attribution after VisEdit edits a counterfactual sample. "BE" and "AE" indicate Before Editing and After Editing respectively. L* indicates the layer index.
• Figure 5: Module contribution of BLIP-OPT (2.7B) and MiniGPT-4 (7B).